[UnifiedRadixTree]: Add KL accuracy CI for UnifiedTree with HiCache (#24346)

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -1244,12 +1244,6 @@ class UnifiedRadixCache(BasePrefixCache):
                 self.dec_lock_ref(ancestor_node)
                 return None
 
-        logger.info(
-            "load_back: kv_tokens=%d, node_id=%d",
-            kv_tokens,
-            last_hit_node.id,
-        )
-
         # Load H→D
         aux_xfers = [x for xfers in comp_xfers.values() for x in xfers]
         aux_xfers.extend(anchor_kv_shared_indices_xfers)
@@ -1367,14 +1361,9 @@ class UnifiedRadixCache(BasePrefixCache):
         req = params.req
 
         if last_node.evicted or params.host_hit_length > 0:
-            logger.info(
-                "init_load_back triggered: node_id=%d, host_hit_length=%d",
-                last_node.id,
-                params.host_hit_length,
-            )
             loading_values = self.load_back(last_node, mem_quota, req=req)
             if loading_values is not None:
-                logger.info(
+                logger.debug(
                     "init_load_back success: loaded %d tokens for node %d",
                     len(loading_values),
                     last_node.id,

python/sglang/test/kl_test_utils.py

@@ -118,8 +118,13 @@ def compare_kl_divergence(
 
 
 # Common request helpers
-def _flush_cache(base_url):
-    requests.post(base_url + "/flush_cache")
+def _flush_cache(base_url, timeout_s=30):
+    response = requests.post(
+        base_url + "/flush_cache",
+        params={"timeout": timeout_s},
+        timeout=timeout_s + 10,
+    )
+    response.raise_for_status()
 
 
 def _generate(